Hugging Face AI Updates: September 26, 2026

1. huggingface_hub 2.0 Moves to httpx2 and Removes Every 1.x Deprecation

Hugging Face shipped huggingface_hub v2.0.0, a breaking release that replaces the httpx HTTP stack with httpx2, so custom client factories and exception handlers must switch types, and TLS now uses the OS certificate store by default. It also deletes every API deprecated in 1.x, including upload_large_folder, duplicate_space, and the old huggingface-cli entry point in favor of hf. Code that needs to support both major versions can import the HTTP module from huggingface_hub.utils, and a migration guide covers the rest. Source

2. TRL 1.14 Dropped Its Fused Loss Copies After They Silently Broke DDP Gradients

Hugging Face released TRL v1.14.0, which removes the trl.losses module and has DPO, KTO, and GRPO stream per-token log-probs into their existing loss code instead of reimplemented fused losses. The team found the duplicate formulas had drifted, including a plain-DDP path where gradients were never all-reduced across ranks. The change unlocks precomputed reference log-probs with Liger, cutting step time 12.1% and peak memory 31% on a Qwen3-0.6B H100 test, and a new in-tree fused Triton logprob and entropy kernel runs about 14x faster than the PyTorch path. Source